328
Volume 60 – Number 2
PROTECTING VIRTUAL THINGS: PATENTABILITY OF ARTIFICIAL
INTELLIGENCE TECHNOLOGY FOR THE INTERNET OF THINGS
ANASTASIA GREENBERG*
ABSTRACT
The Internet of Things (IoT) welcomes physical
everyday objects into the connected digital world. Existing
IoT devices include virtual assistants, smart thermostats,
and fitness trackers. IoT innovation is booming, bringing
to market increasingly sophisticated devices with immense
potential for improving human well-being in the areas of
smart cities, efficient manufacturing, and personalized
healthcare. The true engine behind the IoT revolution is
artificial intelligence (AI) which uses computing power to
learn from big data generated by IoT sensors to deliver
smart solutions and accurate predictions—furnishing IoT
devices with their value. To gain a bird’s eye perspective
on the future development of AI applications for IoT
(termed AI-IoT in this Article), one important consideration
is whether such technology can enjoy intellectual property
protection in the form of patents, and what the
consequences are of such patents on the AI-IoT innovation
landscape. Part I of this Article introduces the concepts of
AI and machine learning and describes criteria for
* Ph.D.; Associate in Intellectual Property, WilmerHale, Boston. I
graduated from McGill University’s Faculty of Law and hold a Ph.D. in
Neuroscience from the University of Alberta. The content of this
Article is my own view, was developed and submitted for publication
while I was enrolled at McGill University’s Faculty of Law, and does
not represent the views of any of the organizations that I am currently
affiliated with. Correspondence: [email protected].
Protecting Virtual Things 329
Volume 60 – Number 2
obtaining a patent under United States intellectual property
law. Part II of this Article covers the historical
background of the subject matter eligibility of software
patents through jurisprudential and policy developments
with a focus on implications for the patentability of AI-IoT.
Part III of this Article addresses innovation policy
consequences of proliferation of AI-IoT patents. The
Article finds that AI-IoT patents present a unique set of
tangible inventions that may circumvent the “abstract
idea” obstacle to subject matter eligibility faced by many
software patents. However, current evidence is ambiguous
as to whether the growth of such patents would stimulate or
dampen AI-IoT innovation. In any case, AI-IoT patents
should be welcomed as current patent law does not have a
clear legal test for exempting such patents and a
technologically-neutral approach to intellectual property
should be embraced.
Abstract ........................................................................... 328
I. Introduction ............................................................. 330
A. Artificial Intelligence for IoT
Devices (AI-IoT) ................................................. 331
B. Artificial Intelligence and Machine Learning ..... 333
C. What is a Patent? ................................................. 335
II. Are AI-IoT Inventions Patentable Subject
Matter? .................................................................... 336
A. The Alice Decision ............................................. 338
B. The Aftermath of the Alice Decision .................. 339
C. Patentability Challenge and Promise
for AI-IoT............................................................ 340
330 IDEA – The Law Review of the Franklin Pierce Center for Intellectual Property
60 IDEA 328 (2020)
III. Innovation Policy Perspective on
Patenting AI-IoT ..................................................... 344
A. Evidence for Negative Impact of
Software/AI Patents on Innovation ..................... 345
B. Evidence for the Positive Impact of
Software/AI Patents on Innovation ..................... 348
IV. Conclusion .............................................................. 349
I. INTRODUCTION
The Internet gave rise to a virtual world, and the
Internet of Things (IoT) will merge that virtual world with
the physical one. IoT consists of networks of physical
devices connected to the Internet which gather data from
their environment using sensors, share information across
the network, and allow for intelligent data analysis.1
Digitization of the physical world through IoT is expected
to drive the fourth industrial revolution.2 Bain estimated
that the global IoT market will grow from $235 billion in
2017 to $520 billion by 2021.3 The main areas of
application of IoT technology include smart cities for
managing traffic and other public infrastructure,
autonomous vehicles, worksite infrastructure for predictive
1 Amy JC Trappey et al., A Review of Essential Standards and Patent
Landscapes for the Internet of Things: A Key Enabler for Industry 4.0,
33 ADVANCED ENGINEERING INFORMATICS 208, 208 (2017). 2 Jean-Marc Frangos, The Internet of Things will Power the Fourth
Industrial Revolution. Here’s How, WORLD ECON. FORUM (June 24,
2017) https://www.weforum.org/agenda/2017/06/Internet-of-things-
will-power-the-fourth-industrial-revolution [https://perma.cc/Q2C2-U
WV5]. 3 Ann Bosche et al., Unlocking Opportunities in the Internet of Things,
BAIN (Aug. 7, 2018), https://www.bain.com/insights/unlocking-opport
unities-in-the-internet-of-things [https://perma.cc/VN5H-HPBX].
Protecting Virtual Things 331
Volume 60 – Number 2
maintenance, security, precision farming, and connected
health through wearables.4 As a practical example of a
smart city IoT application, the cities of Doha, Sao Paulo,
and Beijing use sensors attached to water infrastructure to
monitor and mitigate water loss.5 Since 2014, there are
more IoT devices in use than the world’s human
population.6
A. Artificial Intelligence for IoT Devices (AI-
IoT)
Despite the significant promise of IoT for both
economic and social benefit, the full potential of IoT
remains unrealized. IoT devices gather massive amounts of
complex data, with only a small portion of that data being
analyzed for practical ends. For example, McKinsey
Global Institute claimed that less than one percent of the
data being collected by thirty thousand sensors on a
specific oil rig are used in decision-making.7 The key to
extracting the maximum value from such “big data” (i.e.,
large datasets with many sources and variables) is through
intelligent data processing and analysis using artificial
intelligence (AI).8 AI is defined as “the capability of a
machine to imitate intelligent human behavior [and
4 LexInnova, Internet of Things: Patent Landscape Analysis, WIPO 1,
2–3 (2014), https://www.wipo.int/edocs/plrdocs/en/internet_of_things.
pdf [https://perma.cc/K4CF-LL48]. 5 Id.
6 See Trappey et al., supra note 1, at 209.
7 James Manyika et al., Unlocking the Potential of the Internet of
Things, MCKINSEY GLOBAL INSTITUTE (June 2015), https://www.
mckinsey.com/business-functions/mckinsey-digital/our-insights/the-int
ernet-of-things-the-value-of-digitizing-the-physical-world [https://perm
a.cc/7DVY-NJTF]. 8 Mohammad Saeid Mahdavinejad et al., Machine Learning for Internet
of Things Data Analysis: A Survey, 4 DIGITAL COMM. & NETWORKS
161, 161 (2017).
332 IDEA – The Law Review of the Franklin Pierce Center for Intellectual Property
60 IDEA 328 (2020)
intuition].”9 AI algorithms are applied to big data to extract
meaning from these data by categorizing information,
finding patterns, and making predictions. To truly
appreciate the power of AI for analyzing big data, it is
important to understand that while AI imitates “human
intuition,” unlike AI, human intuition fails at extracting
relevant patterns from big data and drawing accurate
conclusions based on these patterns. Most AI technology
finds its applications in analyzing big data on the Internet.10
Researchers were able to predict flu trends using data
obtained from Twitter, Facebook uses AI for facial
recognition of users’ image posts, and Netflix uses AI to
make personally catered movie/show recommendations to
subscribers.11
Such AI-based analysis is arguably the most
critical component of IoT. AI is essential for training
autonomous vehicles to make decisions, predicting health
issues from data obtained by wearable devices, and
regulating congestion from traffic data. It is therefore the
combination of IoT and AI that marks the entry point to the
next industrial revolution. This Article examines the
current and future innovation landscape for IoT technology,
with a specific focus on AI software development for IoT
(AI-IoT). Longstanding legal theory suggests that
intellectual property rights are essential for incentivizing
creation by giving creators/inventors a time-limited
monopoly on the fruits of their labor in exchange for public
dissemination of knowledge. Under this framework, the
Article asks whether AI-IoT inventions can enjoy patent
9 Artificial Intelligence, MERRIAM-WEBSTER, https://www.merriam-
webster.com/dictionary/artificial+intelligence [https://perma.cc/24UH-
MR3Z] (last visited Nov. 10, 2019). 10
Hidemichi Fujii & Shunsuke Managi, Trends and Priority Shifts in
Artificial Intelligence Technology Invention: A Global Patent Analysis,
58 ECON. ANALYSIS & POL’Y 60 (2018). 11
Hyunjong Ryan Jin, Think Big! The Need for Patent Rights in the
Era of Big Data and Machine Learning, 7 N.Y.U. J. INTELL. PROP. &
ENT. L. 78, 102 (2018).
Protecting Virtual Things 333
Volume 60 – Number 2
protection under United States’ law. The focus is on
patents since patents protect the functional aspects of the
invention, while copyright protection is concerned with the
literal copying of software code. The Article assumes that
the true value of AI-IoT applications is its technical
function which can be protected through patents. The focus
is also on U.S. law since the majority of both IoT and AI
patents are filed in the US.12
The last part of the Article
turns to policy considerations discussing the advantages
and drawbacks of using patent law for incentivizing
innovation in the AI-IoT space.
B. Artificial Intelligence and Machine
Learning
Before delving into whether AI technologies are
patentable, it is crucial to understand in some detail how AI
algorithms work. The term AI is most often used to refer to
a specific category of algorithms called machine learning
(ML) that allows computers to learn from data without
being explicitly programmed or “hard-coded.”13
ML
algorithms are “trained” on complex data sets and are able
to learn relevant patterns and correlations from
“experience.”14
There are three main categories of ML:
supervised, unsupervised, and reinforcement learning. In
supervised learning, the input data (i.e., the “training data”)
is labeled with the correct response and the algorithm
learns the relationship between the data and the labels to
make predictions on new, previously unseen data.15
An
example of supervised learning is an algorithm that is
trained on many pictures labeled as either containing or not
containing a cat, to then be able to identify a cat picture that
12
Trappey et al., supra note 1, at 219. 13
Mahdavinejad et al., supra note 8, at 165. 14
Jin, supra note 11, at 88. 15
Jin, supra note 11, at 89.
334 IDEA – The Law Review of the Franklin Pierce Center for Intellectual Property
60 IDEA 328 (2020)
it has not previously seen. In unsupervised learning, the
algorithm is fed a complex data set, but without any labels.
The algorithm finds interesting patterns in the data without
being shown any correct solutions. An example of this is
an algorithm that is given a compilation of news articles
and the algorithm learns to group all the articles about the
same news event into one cluster.16
In the context of AI-
IoT, an algorithm fed heart rhythm data from a wearable
could be trained to recognize abnormal heart activity either
through supervised learning by being shown previously
labeled examples of abnormal heart signals, or through
unsupervised methods by using the data to categorize
different heart activity patterns into groups without labels
(a person will then decide which group contains the
abnormal heart rhythms). Reinforcement learning involves
algorithms learning sequences of actions to be taken for a
given situation in order to maximize payoff, such as
training a robot to make a series of complex decisions when
playing soccer.17
The steps involved in developing a ML algorithm
are (1) obtaining high-quality data for training the
algorithm, such as data acquired by sensors on IoT devices;
(2) pre-processing data including cleaning data by
removing outliers or reducing dimensions; (3) training a
ML algorithm on the data (the ML algorithm is either an
off-the-shelf algorithm commonly used or a newly
developed one); and (4) obtaining a final trained algorithm
(i.e., the model) which gives output data (solutions) when
shown new input data.18
While it is true that ML
algorithms learn from the data and come up with a final
model spontaneously, the developer’s ingenuity still plays a
major role. Many human decisions need to be made during
16
Jin, supra note 11, at 89. 17
Mahdavinejad et al., supra note 8, at 165. 18
Jin, supra note 11, at 92.
Protecting Virtual Things 335
Volume 60 – Number 2
the development process including choosing which ML
method(s) to employ for a given problem, how to curate the
training data, which algorithm parameters to select, and
how to test the model for accuracy. When considering a
patent for an AI/ML invention, an inventor may seek
protection for either a single development step, or more
commonly, a series of these steps presented as a whole.
C. What is a Patent?
Section 101 of the Patent Act allows for four types
of inventions to receive patent protection: (1) processes; (2)
machines; (3) manufactures; and (4) compositions of
matter.19
AI-IoT would fall under either process (e.g., steps
in algorithm implementation) or machine (e.g., AI
combined with a physical IoT device). The term of a patent
is twenty years from the date on which the application for
the patent was filed in the United States.20
In order to
obtain a patent under one of these categories, an invention
needs to meet all of the following criteria: it must be (1)
patent-eligible subject matter; (2) useful; (3) novel; and (4)
non-obvious.21
The courts have interpreted patentable
subject matter to mean that almost any invention is eligible
except for laws of nature, natural phenomena, and abstract
ideas.22
Useful means that the invention has a practical
application, not merely a theoretical application in the
future. Novel means that the invention does not repeat
“prior art” (i.e., previously patented inventions), has not
been publicly disclosed, and is not generally known. Non-
19
35 U.S.C. § 101 (2018). 20
35 U.S.C. § 154(a)(2) (2018). 21
35 U.S.C. § 101 (patentable subject matter and utility); 35 U.S.C. §
102(a) (2018) (novelty); 35 U.S.C. § 103 (2018) (non-obviousness). 22
Allen Clark Zoracki, When Is an Algorithm Invented? The Need for a
New Paradigm for Evaluating an Algorithm for Intellectual Property
Protection, 15 ALB. L.J. SCI. & TECH. 579, 589 (2004).
336 IDEA – The Law Review of the Franklin Pierce Center for Intellectual Property
60 IDEA 328 (2020)
obvious inventions are those that would not have been
obvious to others “skilled in the art” (i.e., experts in the
field). Therefore, requiring some kind of creative insight
rather than simply combining previous
inventions/knowledge in a standard way is a prerequisite to
being patentable. For AI-related inventions, eligible
subject matter represents the largest hurdle to overcome
since algorithms and mathematical formulae have been
traditionally considered to be abstract ideas.23
II. ARE AI-IOT INVENTIONS PATENTABLE SUBJECT
MATTER?
The U.S. Supreme Court has stated that
“phenomena of nature, though just discovered, mental
processes, and abstract intellectual concepts are not
patentable, as they are the basic tools of scientific and
technological work.”24
The courts are against issuing
patents for such work since “monopolization of those tools
through the grant of a patent might tend to impede
innovation more than it would tend to promote it.”25
Although advanced AI technology, and its use in IoT, is
relatively new, the courts have been grappling with the
patentability of computer software inventions for almost
half a century.
Prior to the 1980s, it was generally accepted that
software represented abstract mathematical concepts and
remained unpatentable subject matter.26
In 1981, the
Supreme Court ruled in Diamond v. Diehr that a formula,
implemented on a digital computer, for curing rubber was
23
Id. at 588. 24
Gottschalk v. Benson, 409 U.S. 63, 67 (1972). 25
See Mayo Collaborative Servs. v. Prometheus Labs., Inc., 132 S. Ct.
1289, 1293 (2012). 26
See Gottschalk, 409 U.S. at 67.
Protecting Virtual Things 337
Volume 60 – Number 2
patentable.27
The patent included the process of “installing
rubber in a press, closing the mold, constantly determining
the temperature of the mold, constantly recalculating the
appropriate cure time through the use of the formula and a
digital computer, and automatically opening the press at the
proper time.”28
Given that the computer software improved
a physical-industrial process as a whole, the Court did not
view the invention as abstract, and therefore it was upheld
as valid.29
Following this decision, the United States Court
of Appeals for the Federal Circuit became more open to
accepting that software can be patent-eligible subject
matter. In Alappat, the Federal Circuit accepted the
validity of a software patent that processed data to allow for
the display of a smooth waveform on a digital
oscilloscope.30
The court viewed the patent as creating a
machine in which a general-purpose computer was turned
into a special-purpose computer when running the software
to digitize the waveforms.31
In State Street Bank and Trust Company v.
Signature Financial Group, Inc., the Federal Circuit found
that an algorithm used to calculate a share price was patent-
eligible since it constituted a practical and tangible
application of a mathematical formula that could be used
for recording and reporting purposes.32
In State Street
Bank, the court also articulated the “machine-or-
transformation” test which provided that an algorithm or
software is patentable if (1) it is tied to a particular machine
or apparatus; or (2) it transforms a particular article into a
27
See Diamond v. Diehr, 450 U.S. 175 (1981). 28
Id. at 187. 29
Id. at 191–92. 30
See In re Alappat, 33 F.3d 1526, 1544 (Fed. Cir. 1994). 31
See id. 32
See State St. Bank & Trust Co. v. Signature Fin. Grp., Inc., 149 F.3d
1368 (Fed. Cir. 1998).
338 IDEA – The Law Review of the Franklin Pierce Center for Intellectual Property
60 IDEA 328 (2020)
different state or thing. Later, in Bilski v. Kappos, the
Supreme Court clarified that the “machine-or-
transformation” test was useful but not the sole criterion for
determining subject matter eligibility.33
Following State
Street Bank, the United States Patent and Trademark Office
(USPTO) saw a proliferation of software- and Internet-
related patent applications.34
The decision opened a can of
worms, more than doubling the annual number of software
patent applications in the years following the decision,
including patent applications implementing basic
calculations on a computer, as well as “business method”
patents.35
The most famously known business method
patent granted following State Street Bank is Amazon’s
“one-click” patent allowing customers to make single-click
purchases based on previously stored payment
information.36
A. The Alice Decision
The Supreme Court did not return to the question of
software subject matter eligibility for almost two decades
until the case of Alice Corp. v. CLS Bank International in
2014, marking a major turning point in the fate of software
patents.37
The patent in question involved a computerized
method of mitigating settlement risk by keeping track of
each party’s account balance to prevent one party from
33
See Bilski v. Kappos, 561 U.S. 593 (2010). 34
Fabio E. Marino & Teri H. P. Nguyen, From Alappat to Alice: The
Evolution of Software Patents, 9 HASTINGS SCI. & TECH. L.J. 1, 6
(2017). 35
Christopher W. Quinn, The 20 Year War On Patents: When Will It
End?, LEXOLOGY https://www.lexology.com/library/detail.aspx?g=8cd
d3dd7-1fb3-48dc-a7ba-b6ffb0076e8e [https://perma.cc/9UZT-UBXY]
(last visited Nov. 10, 2019). 36
U.S. Patent No. 5,960,411. 37
See Alice Corp. v. CLS Bank Int’l, 573 U.S. 208 (2014).
Protecting Virtual Things 339
Volume 60 – Number 2
reneging on the deal.38
The Court developed a two-step
test for determining subject matter eligibility: (1) determine
whether the claims are directed to a patent-ineligible
subject matter, such as an abstract idea; and if so, (2)
determine whether the claims as a whole transform the
patent ineligible subject matter into a patentable invention
through an inventive concept.39
In this case, the Alice
patents failed step one since the invention did nothing more
than implement the abstract idea of mitigating settlement
risk on a generic computer.40
The major concern with the
Alice decision is that the Court failed to provide a definition
of “abstract idea,” leaving the state of software patents in
muddled confusion.41
B. The Aftermath of the Alice Decision
In the aftermath of the Alice decision, lower courts
were left to interpret the meaning of abstract in applying
the two-step framework. A host of post-Alice decisions
began to rely on the “mental steps” doctrine for assessing
patent eligibility which posits that if a software
program/algorithm is performing a process that could be
performed by a person using solely his or her mind or by
using a pen and paper, then the patent will be presumed
ineligible subject matter.42
In contrast, a series of Federal
Circuit decisions pointed toward the willingness to uphold
the validity of software that improved “computer function.”
38
Id. at 212. 39
Id. at 218–21. 40
Id. at 221. 41
Robert Daniel Garza, Software Patents and Pretrial Dismissal Based
on Ineligibility, 24 RICH. J.L. & TECH. 1, ¶ 29 (2017). 42
Ben Hattenbach & Gavin Snyder, Rethinking the Mental Steps
Doctrine and Other Barriers to Patentability of Artificial Intelligence,
19 COLUM. SCI. & TECH. L. REV. 313, 317–18 (2017).
340 IDEA – The Law Review of the Franklin Pierce Center for Intellectual Property
60 IDEA 328 (2020)
In DDR Holdings LLC v. Hotels.com LP, the
invention concerned integrating third-party merchant web
content with the “look and feel” of a host webpage which
prevented the host webpage from losing visitor traffic when
directed to the content of an advertiser.43
The court
decided that this invention was not directed to an abstract
idea under the first step of the Alice test since it solved a
problem in computer networks. Similarly, in McRO Inc. v.
Bandai Namco Games America Inc., the Federal Circuit
upheld a patent for software that automatically
synchronized the facial expressions and lip movements of
three-dimensional animated characters.44
The court
determined that this invention was not abstract because it
improved computer animation technology.45
In attempt to
fill the gaps in the Alice ruling, the courts seem to favor
certain software patents over others, particularly those that
“improve computer technology” in ways distinguished
from human “mental processes.” The next section will
discuss how this current state of software subject matter
eligibility affects the patentability of AI technology for IoT.
C. Patentability Challenge and Promise for
AI-IoT
In response to the Alice decision, the USPTO issued
a guidance document on subject matter eligibility which
defined four broad categories of inventions that represent
“abstract ideas”: (1) those that emulate mental processes;
(2) those that can be replaced with pen and paper; (3) those
that focus on human interaction; and (4) those that solve a
problem which existed before the invention of the
43
See DDR Holdings, LLC v. Hotels.com, L.P., 773 F.3d 1245 (Fed.
Cir. 2014). 44
See McRO, Inc. v. Bandai Namco Games America Inc., 837 F.3d
1299 (Fed. Cir. 2016). 45
Id. at 1316.
Protecting Virtual Things 341
Volume 60 – Number 2
Internet.46
The guidance appears problematic for AI
inventions since AI learns through experience with data
which parallels human learning through “mental
processes,” and attempts to solve a host of problems that
existed before the Internet (e.g., traffic, medical diagnosis,
etc.). The human interaction category would also
potentially exclude patentability of any AI-IoT
technologies that connect to social networks, make phone
calls, and so forth.
The USPTO abstract idea categories are at the very
heart of what AI is designed to do. AI is being developed
to understand human language, identify patterns, and make
accurate predictions with increasing sophistication—tasks
traditionally performed by the human brain. In Blue Spike
LLC v. Google Inc, the patent at issue created an AI-based
method of identifying and comparing digital signals with
high accuracy.47
For example, the algorithm could be fed a
piece of digitally-stored music and identify that it contained
a cover of an original, copyrighted song.48
A California
District Court held that the patent was abstract since it
concerned the mental processes of identifying and
recognizing signals on computers (i.e., a person could listen
to a song and identify it as a cover). Similarly, in
Purepredictive Inc. v. H20.AI Inc., the same California
District Court found that a patent based on AI predictive
analytics is abstract since it covers “mental processes.”49
46
See Interim Guidance on Patent Subject Matter Eligibility, 79 Fed.
Reg. 74,618 (Dec. 16, 2014). 47
See Blue Spike, LLC v. Google Inc., No. 14-cv-01650-YGR, 2015
U.S. Dist. LEXIS 119382 (N.D. Cal. Sept. 8, 2015). 48
Id. at *2–5. 49
See Purepredictive, Inc. v. H2O.AI, Inc., No. 17-cv-03049-WHO,
2017 U.S. Dist. LEXIS 139056 (N.D. Cal. Aug. 29, 2017) at *13.
342 IDEA – The Law Review of the Franklin Pierce Center for Intellectual Property
60 IDEA 328 (2020)
The USPTO guidance and the above-mentioned
court decisions seem to conflate human cognition with AI
technology without clear distinctions, leaving many
unanswered questions about the patentability of AI. While
it is true that AI addresses problems that human cognition
can also be used to address, does the exact method of
solving those problems matter for patent eligibility? What
about the accuracy of solutions produced by AI compared
to humans? While a physician uses cognition to diagnose
tumors in breast biopsy images, an AI system can identify
tumors using complex “black box” patterns and
correlations that are highly unlikely to match the method of
human cognitive processing. AI is also able to arrive at a
larger number of accurate diagnoses.50
The Federal Circuit
seemed to have answered the second question by stating
that “relying on a computer to perform routine tasks more
quickly or more accurately is insufficient to render a claim
patent eligible.”51
The answer to the question of the
method used to arrive at a solution remains unanswered.
With respect to AI that is specific to IoT, there is an
added physical component of the invention relating to the
IoT device itself. In this case, the “machine-or-
transformation” test from State Street Bank would perhaps
lean in favor of finding AI-IoT patents as subject matter
eligible and non-abstract.52
In accordance with this view,
patent attorney Vincent Spinella-Mamo said that when
preparing AI-related patent applications he tends to include
tangible sources of data such as “physical sensors, data
50
DAYONG WANG ET AL., DEEP LEARNING FOR IDENTIFYING
METASTATIC BREAST CANCER (2016), https://arxiv.org/pdf/1606.0571
8.pdf [https://perma.cc/Z4MX-CUB8]. 51
See OIP Techs., Inc. v. Amazon.com, Inc., 788 F.3d 1359, 1361
(Fed. Cir. 2015) (emphasis added). 52
State St. Bank, 149 F.3d at 1373.
Protecting Virtual Things 343
Volume 60 – Number 2
derived from physical sensors, etc.”53
Although there is
virtually no established law on the patentability of AI-IoT,
it seems that tying the AI technology to physical devices
provides a more secure path to eligibility.
Despite the legal ambiguity, the USPTO has a
patent classification, class 706, that is exclusively
designated for AI data-processing inventions, which at least
inadvertently admits that AI could be patentable.54
In the
past six years, the USPTO has seen a five hundred percent
increase in the number of patents under class 706.55
Google has filed the largest number of AI patent
applications to date, with Amazon in second place—most
of these patents are in IoT in the areas of self-driving cars,
robotics, delivery drones, AI assistants, and health-related
wearables.56
In April 2018, Andrei Iancu, the director of the
USPTO, spoke before the U.S. Senate Judiciary Committee
on various intellectual property issues including that of
patent subject matter eligibility. His comments were highly
53
Vincent Spinella-Mamo, Patenting Algorithms: IP Case Law and
Claiming Strategies, IP FOLIO, http://blog.ipfolio.com/patenting-
algorithms-ip-case-law-and-claiming-strategies [https://perma.cc/5532-
4HDT] (last visited Nov. 10, 2019). 54
Class 706 Data Processing - Artificial Intelligence, USPTO,
https://www.uspto.gov/web/patents/classification/uspc706/defs706.htm
[https://perma.cc/LBW4-W2L8] (last visited Nov. 10, 2019). 55
Frank A. DeCosta, Intellectual Property Protection for Artificial
Intelligence, FINNEGAN (Aug. 30, 2017), https://www.finnegan.com/en
/insights/intellectual-property-protection-for-artificial-intelligence.html
[https://perma.cc/2C22-VD6A]. 56
Winners and Losers in the Patent and Innovation Wars Between
Amazon, Google, Facebook, Apple, and Microsoft, CB INSIGHTS RES.
(Nov. 16, 2017), https://www.cbinsights.com/research/innovation-paten
ts-apple-google-amazon-facebook-expert-intelligence [https://perma.cc/
GF6C-ZVM8].
344 IDEA – The Law Review of the Franklin Pierce Center for Intellectual Property
60 IDEA 328 (2020)
suggestive of a shift in the USPTO’s perspective on AI
inventions. He stated:
This is one place where I believe courts have
gone off the initial intent. There are human-
made algorithms, human-made algorithms
that are the result of human ingenuity that
are not set from time immemorial and that
are not absolutes, they depend on human
choices. Those are very different from
[abstract mathematical concepts such as]
E=mc2 and they are very different from the
Pythagorean theorem.57
More clarity is needed on how the USPTO, courts,
and legislators intend to interpret subject matter
patentability in the age of AI and IoT. Nevertheless,
Iancu’s recent statements combined with the physical
nature of IoT devices suggests that AI-IoT may
increasingly become recognized as patentable subject
matter, offering some clarity for inventors.
III. INNOVATION POLICY PERSPECTIVE ON
PATENTING AI-IOT
Even if AI-IoT technologies are likely to fall within
the scope of patentable subject matter, is it desirable from
an innovation policy perspective to have a proliferation of
such patents? The ongoing dialogue on subject matter
eligibility appears to be a legal disguise for an innovation
policy-based concern of permitting software and AI
patents. Intellectual property rights are protected in the
57
Statement of Director Andrei Iancu Before the S. Comm. on the
Judiciary, USPTO (Apr. 18, 2018), at 1:07, https://www.judiciary.senat
e.gov/meetings/oversight-of-the-us-patent-and-trademark-office [https:
//perma.cc/4Z8Y-6KE7].
Protecting Virtual Things 345
Volume 60 – Number 2
U.S. Constitution, “to promote the progress of science and
useful arts, by securing for limited times to authors and
inventors the exclusive right to their respective writings and
discoveries.”58
The economic idea behind intellectual
property protection is to incentivize the investment and
risk-taking required for innovation by granting a time-
limited monopoly on an invention.59
However, intellectual
property protection is a double-edged sword as
overprotection can stifle downstream innovation by
increasing the transaction costs associated with building off
of protected inventions. It appears that courts and the
USPTO have been implicitly concerned about the second
possibility for software and AI. Empirical evidence on the
role of AI patents in innovation exists for both sides of the
debate. The following sections focus on AI and software
innovation in general given that little research currently
exists for AI-IoT-specific technologies.
A. Evidence for Negative Impact of
Software/AI Patents on Innovation
A study by Ronald Mann found that software start-
up firms do not engage in a “prior art” search of existing
patents before beginning product development, suggesting
that software patents do not directly promote downstream
innovation.60
Another study using data from over one
thousand AI start-ups found that only twenty-one percent of
AI start-ups applied for a patent, with just eleven percent
being granted at least one patent.61
The larger and more
58
U.S. CONST. art. I, § 8, cl. 8. 59
Zoracki, supra note 22, at 583. 60
Ronald J. Mann, Do Patents Facilitate Financing in the Software
Industry, 83 TEX. L. REV. 961, 1004 (2004). 61
Cortica, Numenta Hold Top Patents in Artificial Intelligence, CB
INSIGHTS RES. (Apr. 27, 2017) https://www.cbinsights.com/research/
top-artificial-intelligence-startup-patent-holders [https://perma.cc/B6V
Y-JRB5].
346 IDEA – The Law Review of the Franklin Pierce Center for Intellectual Property
60 IDEA 328 (2020)
established technology companies are dominating the
software and AI patent landscape, but not necessarily for
the purpose of promoting innovation.62
The head of the
patent department of Cisco made the following statement to
the Federal Trade Commission: “The time and money we
spend on patent filings, prosecution, and maintenance,
litigation and licensing could be better spent on product
development and research leading to more innovation. But
we are filing hundreds of patents each year for reasons
unrelated to promoting or protecting innovation.”63
Cisco’s
statement points to the idea that large companies are
seeking software patents for defensive reasons (i.e., to
avoid being accused of infringement) rather than as a
reward for innovation. For these same reasons, the League
for Programming Freedom has stood in opposition to
patenting software since the 1990s.64
One proposed reason for why software patents may
stifle innovation is that software innovations are highly
incremental, cumulative, and collaborative rather than
competitive.65
This situation in the software industry is in
stark contrast to the pharmaceutical industry with a much
smaller number of players and resulting patentable products
given the massive expense and timeframe of drug
development. On the other hand, collaborative software
development websites such as GitHub and Stackoverflow,
as well as open source deep learning packages such as
62
Id. 63
Statement from Robert Barr to the Fed. Trade Commission, at 677–
78, FOUNDATION FOR A FREE INFORMATION INFRASTRUCTURE (Feb. 28,
2002) http://swpat.ffii.org [https://perma.cc/LMU9-LRVE]. 64
Against Software Patents, THE LEAGUE FOR PROGRAMMING
FREEDOM (Feb. 28, 1991), https://groups.csail.mit.edu/mac/projects/lpf
/Patents/against-software-patents.html [https://perma.cc/4PHM-CVBJ]. 65
Anton Hughes, Avoiding the Software Patent Problem: An
Alternative Fix for TRIPS Junkies, 14 MURDOCH U. ELECTRONIC J.L.
100, 105 (2007).
Protecting Virtual Things 347
Volume 60 – Number 2
TensorFlow and PyTorch, make AI development
incremental and accessible to any programmer with a
general computer at hand.66
Moreover, a twenty-year
patent exclusivity period can be considered too long in the
fast-paced software industry where a product lifecycle is
around three to five years.67
Although currently
unexplored, it is possible that AI-IoT may fall closer to the
laborious drug development process than the rapid
incremental classical AI development process given that
AI-IoT includes the additional hurdle of integrating the AI
invention with a physical device. If that were the case,
perhaps AI-IoT patents would be more likely to promote
innovation without overprotecting.
Another issue that tips the balance between
promoting and stifling innovation is how narrow or broad a
patent is construed. If a patent is too narrow, covering a
very specific AI invention, it will fail to adequately protect
the invention, and others can easily circumvent
infringement with slight modifications. This situation
would theoretically interfere with the incentive to innovate
given that the afforded protection would be negligible. On
the other hand, if the patent is too broad, it will restrict
downstream innovation by creating a monopoly over a
large area of AI development. The issue of overly broad
patents is particularly pervasive in software and Internet-
related fields.68
A subsidiary of Alphabet, DeepMind,
recently filed patents titled “Generating Video Frames
Using Neural Networks” and “Reinforcement Learning
66
Software Patents are Obsolete in the Age of AI, BHARATH
RAMSUNDAR (June 10, 2017), http://rbharath.github.io/software-
patents-are-obsolete-in-the-age-of-ai [https://perma.cc/APQ8-7Z6T]. 67
Hughes, supra note 65, at 105. 68
Zoracki, supra note 22, at 586.
348 IDEA – The Law Review of the Franklin Pierce Center for Intellectual Property
60 IDEA 328 (2020)
Systems” which appear to be extremely broad.69
While it is
not yet clear if these particular patents will be granted,
Google has successfully obtained a patent on a broad
machine-learning technique called Dropout which is a
general method for addressing overfitting in a neural
network (overfitting is an issue for all machine-learning
development).70
B. Evidence for the Positive Impact of
Software/AI Patents on Innovation
When it comes to measuring innovation through
software start-up financing and long-term success, a study
by Mann found that obtaining patents was positively
correlated with the number of successful venture capital
financing rounds, total investment amount, and longevity of
the company.71
The 2008 Berkeley Patent Survey found
that while only one-third of software entrepreneurs filed for
patents, the majority of venture-backed software start-ups
had obtained patents.72
Admittedly, these findings are
correlational and it is not clear whether patents themselves
contribute to success or whether it is merely the expectation
of investors for start-ups to obtain patents.
69
Mike James, Google’s DeepMind Files AI Patents, I PROGRAMMER
(June 11, 2018), https://www.i-programmer.info/news/105-artificial-
intelligence/11884-googles-deepmind-files-ai-patents.html [https://per
ma.cc/BN4C-E8NT]. 70
U.S. Patent No. 9,406,017. 71
Ronald J. Mann & Thomas W. Sager, Patents, Venture Capital, and
Software Start-Ups, 36 Res. Pol’y 193, 200–03 (2007), https://scholar
ship.law.columbia.edu/faculty_scholarship/1380 [https://perma.cc/M7J
S-NXKN]. 72
Stuart J.H. Graham et al., High Technology Entrepreneurs and the
Patent System: Results of the 2008 Berkeley Patent Survey, 24
BERKLEY TECH. L.J. 1255, 1277 tbl.1 (2009).
Protecting Virtual Things 349
Volume 60 – Number 2
Another important factor in understanding the
impact of patents on downstream innovation is the full
lifecycle of a patent which includes its litigation history and
its sale and licensing history. While litigation can be seen
as a net negative for innovation given its high transactional
cost, sale and licensing activities can be seen as a net
positive since they represent the transfer and dissemination
of knowledge. A study by Chien using a limited dataset of
patents with publicly available historical information,
demonstrated that a software patent is much more likely to
be sold than it is to be litigated over its lifetime.73
On the
issue of licensing, the vast majority of agreements (eighty-
eight percent) included not only patent rights but also the
additional exchange of various trade-secrets, know-how,
source code, bug-fixing guidance, and other proprietary
information.74
Taken together, Chien’s study implies that
software patents are a net positive for innovation,
supporting the transfer of knowledge through the sale and
licensing of patent rights, including the flow of additional
intellectual property in licensing agreements not directly
connected to the patent itself. In summary, evidence points
in both directions when considering how software/AI
inventions impact downstream innovation over time.
IV. CONCLUSION
Considering the mixed and incomplete evidence
from software/AI patents, it is difficult to predict whether
the expansion of AI-IoT patents will enhance or hamper
fruitful progress in the AI/IoT space. Regardless, this
Article puts forward the proposition that restricting AI-IoT
patents based on subject matter ineligibility does not have a
clear and unambiguous legal basis at the current time.
73
Colleen V. Chien, Software Patents as a Currency, Not Tax, On
Innovation, 31 BERKELEY TECH. L.J. 1669, 1700 (2016). 74
Id. at 1715–19.
350 IDEA – The Law Review of the Franklin Pierce Center for Intellectual Property
60 IDEA 328 (2020)
Courts and the USPTO have gone back and forth in trying
to establish clear tests for subject matter eligibility, and the
undefined concept of abstract ideas has made matters
increasingly complicated for understanding AI
patentability. The lack of legal clarity can create barriers to
entry for emerging AI-IoT innovators.
From a pragmatic point of view, it is therefore
recommended that AI-IoT patents be evaluated under the
other eligibility criteria (i.e., utility, novelty, non-
obviousness). It is also imperative to approach patent
applications with a technical understanding of the
underlying technology since the abstract idea concept
seems to misunderstand the differences between AI
technology and human brain function, as well as the overall
advantage of AI. As well, it is imperative to consider
patentability in a technologically-neutral manner. The
abstract idea criteria are clearly biased against AI
technology. Intellectual property law was never developed
to reward or deny protection based on specific forms of
technology, but rather promises to grant protection for
inventions that meet eligibility criteria regardless of the
technology at hand. This is an especially important
consideration in the digital age given that new
technological forms are increasingly unpredictable and
unprecedented.
As the AI-IoT industry matures, further empirical
evaluation on the relationship between patents and
innovation can be assessed. Depending on the consensus
from such investigation, legislators and policymakers can
step in to provide clear and consistent legal and policy rules
on issues of patentability without technological bias to
ensure continued promotion of AI-IoT development. AI-
IoT innovation holds much promise across a swath of
applications from ecological conservation to industrial
Protecting Virtual Things 351
Volume 60 – Number 2
efficiency to promotion of human health. A careful
application of intellectual property law is a key piece in
realizing that potential.